WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

Evaluation of Dynamic Bandwidth Allocation Algorithms in GPON Networks

Sylvia Smolorz Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing Nokia Siemens Networks GmbH & Co. KG Technical University of Denmark Munich DTU Fotonik Germany Kgs. Lyngby sylvia.smolorz@nsn.com Denmark ozimki@gmail.com, {srru,ladit,hewe}@fotonik.dtu.dk Abstract: In this paper, two approaches for Dynamic Bandwidth Allocation in GPON networks are proposed, and validated through simulations in the OPNET modeler. One approach address a Status Reporting scheme, where the bandwidth allocation originates from the client request. The second use a centralized Non Status Reporting scheme. Furthermore, parameters to cope with variances in the trafc pattern is quantied. The results on performance, scalability and efciency show that Status Reporting is utilizing the bandwidth more efcient while the Non Status Reporting provides better QoS for real time services. KeyWords: GPON, dynamic bandwidth allocation (DBA), simulation, status reporting, OPNET

1 Introduction
Bandwidth requirement for providing new services is increasing. Moreover, different types of users have varying needs regarding the amount of bandwidth and transmission delays. Network providers are forced to think about new mechanisms that will distribute the bandwidth among the users and provide high network reliability. That leads to the increased interest in the optical networks suitable for Fiber to the Home (FTTH) and Fiber to the Building (FTTB) solutions. Creation of the networks in which each connected user obtains high QoS, despite the variation of payment is too costly solution. This leads to high requirement for implementing mechanisms that will be responsible for the appropriate sharing of available resources. One of the access network technologies introducing high throughput, low delays and advanced bandwidth control is Gigabit Passive Optical network (GPON) [1]. GPON is designed to transport Ethernet packets over the optical medium using the GPON Encapsulation Method (GEM) [3]. The physical link is fragmented into GEM frames as specied in [1] [3]. Each downlink frame contains a Bandwidth map (BWMap), with information about allowed transmission times for each Optical Network Unit (ONU) for the future uplink frame. Dynamic Bandwidth Assignment (DBA) algorithm is used to calculate the total bandwidth assigned for end nodes [3]. This paper introduces two different DBA algorithms designed for GPON, namely, status and non

status reporting DBA [3] [4], respectively. DBA for GPON networks is not given in ITU-T GPON specications, and hence often proprietary algorithms are used in OLT equipment. This paper provides a proposal for the DBA algorithms, together with the analysis of the impact of available congurations on the network performance, which has not been heavily researched until now. The research focuses on the network performance, its optimum conguration parameters and QoS that can be delivered to the subscribers with different trafc priorities. GPON network modelling and simulation has been performed using OPNET environment [15]. Related work to PON networks has been carried out in [1014] The remainder of this paper is organized as follows: Section 2 provides background knowledge of DBA and trafc types. In section 3, a DBA algorithm is proposed. In section 4, the simulation scenario is described and simulation results are presented in section 5. Section 6 concludes the paper.

2 Dynamic Bandwidth Assignment (DBA)

GPON uses point-to-multi-point connections between central Optical Line Termination (OLT), coordinating network resources and Optical Network Units (ONU) located near the end users. Maximum allowable distance is 20km. Varying raw transmission rates

ISSN: 1109-2734

111

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

are available:1.24416 Gbit/s uplink, 2.48832 Gbit/s downlink for asymmetric services and 2.48832 Gbit/s uplink and downlink for symmetric services. Due to the high available bandwidth, its allocation is based on the Service Level Agreements (SLA) where Quality of Service (QoS) can be granted according to the demand. GPON is using transmission container (T-CONT) mechanism for provisioning differentiated QoS. T-CONT is the logical connection between OLT and ONU, where multiple T-CONT types can be allocated in one ONU. GPON contains ve different TCONTS ouf of which types 2, 3 and 4 that were feasible for dynamic bandwidth allocation were evaluated: 1. T-CONT type 1 for xed rate trafc sensitive to jitter and delay. 2. T-CONT type 2 for on-off type trafc with well dened bitrate and strict delay requirements is provisioned with assured bandwidth. This bandwidth has to be granted to the T-CONTs trafc, if requested. If not used, bandwidth can be reallocated to other T-CONTs, providing that it is available as soon as T-CONT type 2 requires it. 3. T-CONT type 3 is provisioned with assured bandwidth and additionally, it can be granted non-assured bandwidth if all available assured bandwidth is utilized. It is suitable for variable rate, bursty trafc with requirements for average rate guarantee. 4. T-CONT type 4 has no bandwidth guarantee but it has eligibility in best effort bandwidth sharing. It is suitable for variable rate, bursty trafc with no delay sensitivity. 5. T-CONT type 5 is consolidation of other TCONTs, provisioned with both xed and assured bandwidth. Collisions in such high throughput networks are very costly. In order to avoid collisions the OLT allocates upstream transmission intervals per T-CONT in a TDMA fashion. Transmission is coordinated using GPON Transmission Convergence (GTC) frames in both uplink and downlink direction. In downstream frame, OLT transmits BWMap containing timing information indicating when each T-CONT is allowed to transmit data during future upstream frame. The DBA automatically adjusts bandwidth grants to the needs of a particular T-CONT. The DBA uses the T-CONTs activity status as an input to the scheduler. This activity status can be obtain either explicitly through T-CONT buffer status reporting (SR), or implicitly through transmission of the idle GEM frames

when T-CONT does not have enough data to transmit during all granted upstream allocation intervals. The implicit method is referred to as non status reporting (NSR). Assured bandwidth is granted regardless of the overall trafc load. Additional non-assured and best effort bandwidth allocation depends on the available upstream capacity remaining after allocation of the granted bandwidth. Non-assured and Best effort bandwidth can be overbooked, while this can never be the case for assured bandwidth. During bandwidth allocation the following prioritization is applied: Fixed trafc (highest priority) Assured trafc Non-assured trafc Best effort trafc (lowest priority) For both SR and NSR DBA, the OLT traces the activity status of each T-CONT during one DBA cycle. Obtained information becomes input to the scheduler, which thereby allocates transmission opportunities for the next DBA cycle.

3 Proposed Dynamic Bandwidth Assignment Algorithms

3.1 Non Status Reporting DBA
In the NSR algorithm, the OLT estimates bandwidth t allocation for the next DBA cycle Ba (c), required by each T-CONT t, on the basis of the bandwidth usage t during the previous DBA cycle Bu (c 1). At the beginning of each cycle c, at rst the OLT assigns assured bandwidth for each T-CONT t t (Ba Asr), based on the amount of data transmitted in the previous cycle. After this assignment, remaining GPON bandwidth is divided between non ast sured trafc (Ba N asr). In the end the rest of available bandwidth is distributed for the best effort traft c (Ba BE). The assured bandwidth granted for Tt CONTs of type 2 and 3 for the new cycle Ba Asr(c) is t based on the activity in the previous cycle Bu (c 1), as shown in eq. 1. The expansion factor (EF ) is used to provide fast response for variation in trafc, and to ensure that assured bandwidth is always allocated when required by a T-CONT.
t M axasr M int Ba Asr (c) = t u Bu (c 1) EF
t t if Bu (c 1) = M axt Asr t if Bu (c 1) M int Asr otherwise (1)

t Equation 1 veries that Ba Asr(c) does not exceed the maximum allowed bandwidth M axt Asr for TCONT t and is never lower than the minimum bandwidth M int . Asr

ISSN: 1109-2734

112

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

Assignment of additional non-assured grants t Ba N asr(c) eligible for T-CONTs of type 3 depends both on bandwidth used by the T-CONT in the previt ous cycle Bu (c 1) and assured bandwidth already t assigned for current cycle Ba Asr(c) (eq. 2). It is limited by the total allowed bandwidth M axt otal , and no T minimum bandwidth assignment is guaranteed.
t Ba N asr(c) =

for each T-CONT may occur at most every third frame in order to maximize throughput, since small bandwidth chunks lead to high header overhead and packet fragmentation. Allocations are served in a cyclic manner according to eq. 8 and starting from the T-CONT that was last granted bandwidth.
Bf rame N asr t (c) = Ba N asr t (c) frames since last alloc frames per DBA cycle (8)

t t M axtotal M axAsr t Bu (c 1) EF M axt Asr 0

t if Bu (c 1) EF > M axt total t if Bu (c 1) > M axt Asr otherwise (2)

3.2 Status Reporting DBA

In the SR scheduling, the OLT requests the buffer occupancy status from each T-CONT indicating the number of bytes waiting for transmission. At the beginning of a new frame, one T-CONT of each type gets a token. The T-CONT possessing a token is granted allocation slots according to the request, providing that it does not exceed its maximum bandwidth allowed for current cycle. At rst, requests for assured allocations are served. Later, bandwidth allocations are given to consecutive non-assured bandwidth requests. In the end, best effort requests are served. After each T-CONT is served, the next in order T-CONT receives the token. At the beginning of a new frame, T-CONTs with tokens received in the previous uplink frames are served rst.

When the non-assured bandwidth is overbooked, nonassured bandwidth assignments are scaled proportionally to t into the total link capacity C for each TCONT (eq. 3)
if : ( than : Ba N asr (c) =
t
t t Ba Nasrt(c)(C t Ba Asr(c)) t Ba Nasr(c) t

t Ba Asr(c) +

t Ba N asr(c)) > C

(3)

Best effort bandwidth is assigned proportionally to t Bu Be(c 1) scaled by EF (eq. 4), (eq. 6) and limited by M axtotal . OLT should try to grant to each T-CONT minimum amount of bandwidth in each cycle to avoid T-CONT deadlock situation, where ONU has no means of informing OLT about its bandwidth requirements. In highly overbooked networks zero bandwidth allocation in the previous cycle indicates that the link was too congested to serve the best eft fort trafc. Therefore Ba Be(c) is set to innity, and scaled equally for all T-CONTs of type 4 (eq. 5).
t M axtotal Ba Be (c) = t Bu (c 1) EF
t t if Bu (c 1) = M axt total t if Bu (c 1) = 0 otherwise (4)

4 Simulations
Series of simulations were made using OPNET [15], to evaluate the efciency of SR and NSR DBA algorithms for varying GPON network scenarios. All GPON physical layer properties [5] together with framing and protocol overheads were accurately modeled. The transmission timing highly affects the results, especially on the SR algorithm, where negotia-

if : (

Ba Asr t (c) +

Ba N asr t (c) +

Ba Bet (c)) > C

then : Ba N asr t (c) =

t

Ba Bet (c)(C

Ba Asr t (c)+ Ba Be(c)

Ba Nasr t (c))

(5)
t Ba (c)

B t (c) <> t a Bu (c 1)

(6)

When the DBA cycle ends, the scheduler is provided with assignments for a new cycle (eq.1 to eq.5) and creates a bandwidth map for each frame in the new DBA cycle. Assured trafc is allocated for each TCONT once per millisecond, thereby once in every 8 upstream frames(eq. 7).
EF = ceil(
t Ba Asr(c) 8 ) frames per DBA cycle

(7)

Non-assured and best effort trafc grants are not served in strictly regular intervals. Bandwidth grant

Figure 1: Simulated network.

ISSN: 1109-2734

113

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

tion between OLT and ONU has to be performed before calculation of bandwidth grants. All simulations were performed using GPON 2.48832 Gbps links. No bandwidth is reserved for GPON network control and maintenance, since bandwidth requirement for these control channels is too small to have an effect on the simulation results. Each ONU has three T-CONTs: type 2, type 3 and type 4. Each T-CONT of same type has the same bandwidth parameters (maximum and minimum bandwidth, source bandwidth). The packet sizes vary for different scenarios, depending on the trafc load generated per workstation. The source load is generated with the normal distribution of variance equal to the average packet size. In the rst part of simulations (A) optimum GPON DBA parameters were determined for SR frequency request, EF and DBA cycle (scenarios A.1, A.2 and A.3 correspondingly). In every case a trade off was to be made, since improvement on one network parameter often results in degradation of the other (increase of bandwidth utilization, results in an increased transmission delay). Second part (B) provides Fiber to the Building (FTTB) simulation scenarios for observing network performance and trafc QoS for different network congurations. The modelled FTTB topology is shown in in gure 1. The network consists of one OLT and 48 ofce networks, by which two line cards of OLT are fully utilized. Furthermore, each ofce network consists of one ONU and six workstations, spaced within 200m from centered ONU node. Two workstations are used to model trafc load of one TCONT. Each workstation randomly takes a workstation of the same type as a destination. The workstation can be from any network, and one workstation cannot be chosen more than once. The distance between OLT and any of the ONUs is maximum allowed GPON distance of 20 km. FTTB simulation consists of three scenarios: Scenario B.1 was carried out to compare the SR and NSR DBA methods, in the FTTB network with low trafc load. Since bandwidth utilization is not important in such a case, it is the trafc delay (especially uplink delay), that has to be veried. Scenario B.2 models FTTB network, highly utilized with data trafc and with majority of the available bandwidth reserved for the assured trafc of T-CONTs of type 2 and 3. Scenario B.3 is a nearly overbooked network with high utilization by best effort trafc and relatively low assured trafc load.

The following parameters are used to dene the network conguration: AssuredBW - Amount of bandwidth OLT has reserved for a T-CONT. This bandwidth will be given to that T-CONT if required. Otherwise it will be granted to other T-CONTs SourceBW - Data load generated by one TCONT. This bandwidth is generated by one or more workstations connected to T-CONTs ONU - Ratio between bandwidth reserved for high priority trafc of T-CONTs of type 2 and 3, and the GPON link datarate - Ratio between total trafc load generated by all T-CONTs and the GPON link data rate
SourceBW GP ONBW AssuredBW GP ONBW

5 Results
5.1 Network Tuning
Three different sets of simulations were performed to nd the most optimum values for SR request frequency, EF and DBA cycle length. Status Report request frequency SRrF req indicates how often OLT sends SR request to each TCONT. If T-CONT is polled for SR too often, it decreases overal system throughput and may lead to incompatibilities between OLT and ONU, due to incorrect timing. On the other hand too seldom SR poll results in increased transmission delays. Results from network simulation with DBA cycle of 32 ms, SourceBW AssuredBW GP ONBW = 0.63 and GP ONBW = 0.76 with varying SR request frequency is shown in gure 2.

Figure 2: Uplink delay vs. SR frequency It can be seen that the smallest delay experienced by data belonging to each T-CONT type, is observed with SRrF req = 3, which corresponds to
114 Issue 2, Volume 9, February 2010

ISSN: 1109-2734

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

one SR every 3 frames. Lower SRrF req values result in multiple bandwidth assignment for the same data, and correspond to unacceptable high delays. Higher SRrF req does not have signicant inuence on the overall amount of assigned bandwidth as can be seen in gure 3, however it linearly increases the data delay.

Figure 3: Assigned Bandwidth vs. SR frequency The Expansion factor EF determines the ratio for the bandwidth assigned for current DBA cycle c to the trafc used in the previous DBA cycle c 1. High EF results in greater bandwidth assignments for a TCONT. This additional bandwidth compensates for the potential variation in the trafc load. The trade off for this lower delay is less efcient use of GPON link, since often this additional bandwidth is not fully utilized. Results from network simulation with DBA cySource cle of 32 ms, AssuredBW = 0.78 and GP ONBW = 0.7 GP ONBW BW with varying EF is shown in gure 4. Decrease in

never be smaller than or equal to 1, since this would lead to a situation where the bandwidth for future cycle for T-CONTs of type 2 and 3 could either be constant or decrease. This would be an unacceptable case since one cycle with lower data load would lead to reduced bandwidth allocations for future cycles. In performed set of simulations with EF ranging between 1 and 2, 1.25 turned out to be the best value, that would provide reasonable delay for all T-CONT types. Higher EF results in minimal uplink transmission delay decrease for T-CONTs of type 2 and 3, and rapid delay increase for T-CONT of type 4. The optimal observed value is 1.25. DBA cycle length thus determines how often new cyclic bandwidth assignment for all T-CONTs should be recalculated. Results from network simulation with DBA cycle of 32 ms, AssuredBW = 0.78 and GP ONBW SourceBW = 0.7 with varying EF is shown in gure GP ONBW 6. For the short DBA cycle length, T-CONTs of types 2 and 3 suffer higher delays, while T-CONT of type 4 experiences the least delay. The reason for TCONTs of type 2 and 3 suffering higher delays, is the fact that, when trafc load measurements are taken for a shorter period, they are less uniform. With normal distribution over longer time periods, trafc load tends to be the average value of the normal distribution. The

Figure 5: Assigned Bandwidth vs. EF

Figure 4: Uplink Delay vs. EF in link delay of T-CONT types 2 and 3 corresponds to uplink delay increase for T-CONT of type 4, since increased EF provides higher bandwidth allocations for T-CONTs of type 2 and 3, which results in smaller amount of bandwidth left for BE trafc from T-CONT of type 4 as show in gure 5. Expansion factor can

Figure 6: Uplink Delay vs. DBA cycle

ISSN: 1109-2734

115

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

shorter the measuring period, the higher variation can be observed. These delay values are not exceeding 5 ms since the cycle length is short and bandwidth assigned per cycle can be relatively quickly recalculated for new frame. Low DBA cycle time corresponds to more varying trafc assignments per cycle, while for larger cycle times they are more stable. For higher DBA cycle, less bandwidth is assigned per cycle, since during longer periods more information is gathered about T-CONTs load. For T-CONT 4, the short DBA cycle time is better, since all bandwidth remaining after assignments to T-CONTs of type 2 and 3, will be distributed to the T-CONTs of type 4 that have completely utilized trafc assigned in the previous cycle. In order to ensure lowest delays for assured and non-assured trafc DBA cycle should be higher than 26 ms.

5.2 Fiber to the Building (FTTB)

The rst FTTB scenario was simulated with the AssuredBW = 0.56, following conguration: GP ONBW SourceBW GP ONBW = 0.53 and equal source trafc load for all T-CONTs. It provides a comparison of NSR and SR algorithms with relative low trafc load. The results will show mainly the delay due to the scheduling and physical transmission delays, experienced by the transmitted data. Simulation results show that for the low trafc load of scenario B.1, both NSR and SR algorithms serve all T-CONTs well with the SR average uplink delay of 0.6 ms and NSR of 1.6 ms for all T-CONT types. Delay obtained using SR algorithm is lower, since OLT provides almost immediate response for data transmission request. Delay using NSR algorithm is higher since bandwidth for assured and best effort trafc is allocated every 1 ms. Scenario B.2 shows NSR and SR algorithm performance with high trafc load. The following conSource guration was used: AssuredBW = 0.82 GP ONBW = GP ONBW BW 0.89 and equal source trafc load for all T-CONTs. In congested networks, SR is more efciently using network resources. With the SR DBA, transmission delays of 0.65 ms for T-CONTs of type 2 and 3 are similar to the slightly loaded network in B.1. Best effort trafc for T-CONT type 4 also experiences very low delay of 1 ms as shown in gure 7(a). Bandwidth utilization with SR reaches 90% of the total available bandwidth. It can be seen that although T-CONT of type 4 is granted the highest bandwidth it experiences highest delay, due to the fact that bandwidth for best effort trafc is assigned from upstream frame leftovers after assignments for other T-CONTs. For that reason T-CONT of type 4 obtains bandwidth grants less regularly, which results in higher delay. This irreg-

ularity also causes multiple assignment of bandwidth for the same data. T-CONTs of type 4 are asked for SR every third frame. Since priority is given for TCONTs of type 2 and 3, T-CONTs of type 4 usually do not get bandwidth immediately after OLT receives their status reports. Bandwidth for data is assigned by OLT one or two frames after receiving information about waiting data in response to the status report. In the mean time, second SR request is sent and T-CONT informs again about data waiting in the queue. The problem occurs if OLT assigns bandwidth for data before receiving second SR request. In this case OLT gets duplicated information about amount of data waiting in the queue. This leads to the duplicated bandwidth grant, which will most likely not be fully utilized by the T-CONT. Certainly T-CONTs of type 2 and 3 may experience the same duplicated bandwidth assignment in congested network. However T-CONTs of type 4 are more vulnerable to such cases, due to its higher serving irregularity. Duplicated bandwidth assignment could be eliminated by introduction of the status request time stamps. OLT should register the time, when the last status report was obtained from a particular T-CONT. This information could be used in order to verify whether any data assignment has been issued for that T-CONT in the mean time. NSR algorithm is slightly less bandwidth efcient, with data bandwidth utilization of 86%. TCONTs of type 2 and 3 are given more additional but unused bandwidth. As a result not enough bandwidth remains for T-CONTs of type 4. Average of 648 Mbps assigned to T-CONTs of type 4 is not sufcient to serve 744 Mbps source trafc load, as shown in gure 7(a). Considering the fact that again T-CONTs of type 4 often obtain small bandwidth slices left from each frame, their data packets will be heavily fragmented. As a result data throughput for T-CONT of type 4 is very low. 15% of the best effort trafc does not receive bandwidth. In this case transmission delay of that T-CONT type becomes a minor issue. SR DBA proved to be a better choice for this network conguration, since it offers low transmission delays for both real-time and best effort trafc, and enough bandwidth is assigned for all T-CONT types. The third scenario B.3 was chosen to test NSR and SR algorithms with high trafc load, dominated by best effort trafc of T-CONT 4. The following network conguration was used: AssuredBW = 0.3 and GP ONBW SourceBW = 0.9, source load for T-CONTs 2 and 3 GP ONBW - 432 Mbps each and for T-CONT 4 - 1462 Mbps. Results of the simulation are shown in gure 8(a) and 8(b) . The clear advantages and disadvantages of the two DBA algorithms can be observed in this

ISSN: 1109-2734

116

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

(a) Assigned bandwidth

(b) Uplink transmission delay

Figure 7: Simulation results for Scenario B.2.

ISSN: 1109-2734

117

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

(a) Assigned bandwidth

(b) Uplink transmission delay

Figure 8: Simulation results for Scenario 3.

ISSN: 1109-2734

118

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

high utilized network with high percentage of best effort trafc. In this scenario T-CONT of type 2 assured bandwidth is: AssuredBW = 1.1 SourceBW When NSR DBA is used, T-CONTs of type 3 are experiencing smallest delay of average 1.7 ms, while best effort trafc the highest 7.4 ms. T-CONTs of type 3 obtain the best service due to their maximum bandwidth settings. Due to the decrease of safe bandwidth part given to T-CONTs of type 2, they are experiencing slightly higher delay than T-CONTs of type 2. However this average of 3.6 ms delay, is still acceptable, since real-time data requires delay of maximum 5 ms. The maximum bandwidth setting 1.1 times higher than source bandwidth is enough to eliminate the risk of trafc rejection, and it increases the bandwidth utilization to 90%. This utilization is higher than in the previous scenario, where NSR obtained utilization of only 86%. This indicates, that the difference of bandwidth utilization between NSR and SR algorithms is not particularly due to the estimation of bandwidth assignment itself, but mostly due to the additional security bandwidth given to T-CONTs of type 2 and 3. Bandwidth assignment for the T-CONT of type 3 with NSR DBA is much higher than bandwidth assignment for T-CONT of type 3 with SR. This indicates that T-CONTs of type 3 are granted much more bandwidth when EF is used. Despite the reduction of assured bandwidth for T-CONT of type 2, best effort trafc from T-CONTs of type 4 gets bandwidth assignments irregularly, due to high grants for T-CONT 3. In any case, delay below 10 ms for best effort trafc is sufcient. The total data utilization of the network is 90%, both for NSR and SR DBA. Over 5 ms delay for TCONTs of type 2 with SR DBA indicates that this DBA method tends to duplicate bandwidth grants. When new DBA cycle starts and OLT tends to allocate more bandwidth to T-CONT of type 2, due to the congestion in T-CONTs queue. The bandwidth assignment is duplicated from time to time, which leads to reaching maximum bandwidth limit slightly before end of current DBA cycle. Queues get congested, and they are being emptied at the beginning of next cycle, causing duplicated bandwidth assignment. Due to this uneven distribution of bandwidth grants over the cycle in SR DBA, it is necessary to have bandwidth limits slightly higher than the expected source trafc. NSR DBA seems to be better for the real-time trafc, due to the fact that bandwidth assignments are given regularly every 1 ms. Delay of trafc belonging to T-CONTs of type 4 is, on the other hand, lower than when SR DBA is used (2.4 ms comparing to 7.4 ms). The NSR DBA delay for T-CONTs of type 3 remains around 1.7 ms, as in previous scenarios. It can

be observed that throughout the simulations this tends to be the average value for T-CONTs with a total of assured and non-assured trafc grants not less than 1.25 * source bandwidth. This delay seems to be independent of the trafc load of trafc with other priorities, provided that no overbooking for the high priority trafc has been performed. In order to maintain low delays for real-time trafc, when network is being planned, it should be the amount of assured bandwidth granted for T-CONTs of type 2 and 3, and not the source bandwidth itself, which should be included in the calculation of the load. No overbooking can be made for this bandwidth grants.

6 Conclusion
SR algorighm utilizes bandwidth more efciently than NSR. In the low loaded networks where a lot of bandwidth is available, SR algorithm allocates around 5% more bandwidth than T-CONTs source load. Efciency of NSR algorithm depends highly on the expansion factor used, and the assured bandwidth parameter of the T-CONT. If high assured bandwidth is granted, with expansion factor of 1.25, T-CONTs obtain bandwidth grants for up to 129% of the source bandwidth. This security bandwidth allocated for assured trafc is not fully utilized by data. This leads to degradation in throughput for data transmission. Obviously T-CONTs with the lowest priority suffer the highest delay and data drops due to the low bandwidth utilization, since those T-CONTs can only utilize bandwidth chunks left after assignment of assured and non-assured bandwidth grants. In NSR DBA, the compromise of high QoS for T-CONTs with high priority data and T-CONTs with low priority data is the reduction of the assured bandwidth grants. It was veried that the assured bandwidth parameter set to 110% of the source bandwidth is sufcient to keep the average delay below 5 ms. Low average delay is obtained through the periodical bandwidth allocations, with period of 1ms. This implies that selection of the T-CONTs bandwidth parameters should be performed very carefully, since it is possible for the multiple T-CONT types to obtain reasonable service quality. If a generic situation is considered, where the system load may vary, the most reliable solution for efcient handling of delay sensitive trafc is NSR algorithm. This is although for low trafc load and high additional bandwidth, delays, using NSR, are higher than the case when SR is used. Regular assignment of grants every 1 ms gives guarantee for appropriate handling. If the generated high priority trafc does not exceed the available bandwidth and no overbook-

ISSN: 1109-2734

119

Issue 2, Volume 9, February 2010

WSEAS TRANSACTIONS on CIRCUITS and SYSTEMS

Joanna Ozimkiewicz, Sarah Ruepp, Lars Dittmann, Henrik Wessing, Sylvia Smolorz

ing on assured bandwidth is done, data trafc does not experience delay higher than 5 ms. NSR is not vulnerable to the duplicated bandwidth grants, which implies that bandwidth allocated for each T-CONT is used more efciently. Based on the simulation results, it can be concluded that NSR DBA is more reliable in providing agreed QoS, and, with appropriate network conguration optimized for amount of connected ONUs and their trafc requirements, it can sufciently serve all trafc types. The SR algorithm does not provide and guarantee QoS in all network congurations, but it is efciently using available resources. NSR is more recommended for networks, where transmission delay has to be maintained below a specied minimum level. This solution however requires more bandwidth due to the over-allocated bandwidth. The choice of the algorithm should be made depending on the QoS SLAs agreed upon with particular customers and the dominant trafc type in the network. Acknowledgements: Special thanks goes to Omer Alptekin Yurdal for all the support. The research was supported by the Technical University of Denmark (DTU) and Nokia Siemens Networks. References: [1] ITU-T, G.984.1 Gigabit-capable Passive Optical Networks (GPON): General characteristics, Mar. 2003. [2] ITU-T, G.983.2 Broadband optical access systems based on Passive Optical Networks (PON), Jul. 2005. [3] ITU-T, G.984.3. Gigabit-capable Passive Optical Networks (G-PON): Transmission convergence layer specication, Mar. 2008. [4] ITU-T, G. 983.4 A broadband optical access system with increased service capability using dynamic bandwidth assignment, Nov. 2001. [5] ITU-T, G.984.2 Gigabit-capable Passive Optical Networks (GPON): Physical Media Dependent (PMD) layer specication, Mar. 2003. [6] Joanna Ozimkiewicz, Competing Broadband Access Networks, Masters thesis, Danish Technical University, Nokia Siemens Networks, Jan. 2009. [7] OPNET Technologies, Inc., http://www.opnet.com. [8] Aida Salihovic, Matija Ivekovic, Gigabit Passive Optical Network GPON, Jun. 2007. [9] Rich Baca and Muneer Zuhdi, Technological challenges to G-PON operation, Feb. 2008.
ISSN: 1109-2734 120

[10] Joanna Ozimkiewicz and Sarah Ruepp and Lars Dittmann and Henrik Wessing and Sylvia Smolorz, Dynamic Bandwidth Allocation in GPON Networks, Proc. CISST Conference at Harvard University, 2010 [11] M. Thomsen and T. S. Lyster and M. Berger and B. Mortensen and B. Srensen, Development platform for dynamic bandwidth allocation schemes in future MPCP enabled Ethernet Passive Optical Network (EPON),WSEAS Transactions on Communications, 2006, vol. 5, issue 1, pages 92-98 [12] C. G. Park and Y. Lee and D. H. Han and B. Kim, Waiting time analysis of gated polling system for bandwidth allocation in a passive optical network, WSEAS Transactions on Communications, 2005, vol 4 , issue 10, p. 963-970 [13] S. V. Kartalopoulos, Security and bandwidth elasticity aspects of the CWDM/TDM-PON network, WSEAS Transactions on Communications, 2006, vol 5, issue 8, p. 1461-1468 [14] T. Orphanoudakis and H.-C. Leligou and J. D. Angelopoulos, Next generation ethernet access networks: GPON vs. EPON, Proceedings of the 7th WSEAS International Conference on Electronics Hardware, Wireless and Optical Communications (EHAC 08), 2008 [15] OPNET Technologies, OPNET Modeler, www.opnet.com

Issue 2, Volume 9, February 2010